Neural Fine-Grained Sentiment Analysis with Unsupervised and Transfer Learning Approaches
Dr Ng Hwee Tou, Provost'S Chair Professor, School of Computing
09 May 2019 Thursday, 09:00 AM to 10:00 AM
COM1 Level 3
We study the problem of aspect-based sentiment analysis, also referred to as fine-grained sentiment analysis, which is an important area in sentiment analysis and has seen a lot of research effort and real-world applications. Different from document-level and sentence-level sentiment analysis which only assign an overall polarity score to an input text, aspect-level analysis is based on the idea that an opinion should include a sentiment and a target, therefore, it aims to identify the sentiment-target pairs from a given text. For example, the review sentence "Great food but the service is dreadful" evaluates two aspects -- food (positive) and service (negative).
The development of aspect-based sentiment analysis systems generally faces two major challenges. First, this problem is naturally more difficult compared to coarse-grained sentiment analysis because more fine-grained features are needed for aspect-level predictions. Second, the training resources for this task are limited as it is expensive to obtain fine-grained annotated data. Therefore, in this thesis, we focus on two objectives: (1) design flexible and effective models for fine-grained sentiment analysis; (2) leverage unlabeled data or other cheap resources or supervisions with unsupervised and transfer learning approaches.
Specifically, we study two core problems of aspect-based sentiment analysis -- aspect extraction and aspect-dependent sentiment classification. The former aims to extract aspects (e.g. "food" and "service" in the previous example) from opinionated corpus, while the latter aims to predict the sentiments on extracted aspects. In this proposal, we first describe a novel neural attention approach for extracting aspects in an unsupervised learning setting. We demonstrate that the proposed model is able to produce highly meaningful and coherent aspects.
For aspect-dependent sentiment classification, we first propose an effective attention modeling approach to more accurately capture the correct opinion context for each target when multiple sentiment-target pairs appearing in a sentence. We further propose approaches to transfer knowledge from document-level annotated corpora to boost the performance of aspect-dependent sentiment classification.
Finally, we develop an end-to-end solution that simultaneously performs the two fine-grained tasks. It also allows the fine-grained tasks to be trained together with relevant coarse-grained tasks, leveraging the knowledge from larger corpora to alleviate the issue of limited fine-grained training resources. The proposed model enables the informative interactions between different tasks to better exploit the joint information.